{"id":"971a63e0-27b3-4e5f-8251-c4b9f0f749d5","arxiv_id":"2502.03036","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"FuXi-alpha, a sequential recommender with decoupled temporal, positional, and semantic attention channels plus a two-stage FFN, reports gains over HSTU and positive online engagement results.","lead":"This paper introduces FuXi-alpha, a sequential recommendation model that separates temporal, positional, and semantic signals into different attention channels and adds a two-stage feed-forward network. Across public benchmarks and a Huawei Music A/B test, it reports higher accuracy and engagement than existing models, with further gains as the model grows.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The HSTU comparison is not parameter-matched: AMS and MFFN add per-layer parameters, so offline gains and the scaling trend may reflect added capacity, not the proposed mechanism.","rationale":"Independent reading of the paper: the strongest claim is empirical state-of-the-art and scaling behavior, and the released code plus public benchmarks are real supporting evidence. The load-bearing weakness is that the headline comparison with HSTU is not controlled for total parameter count. The reader's weakest assumption identifies exactly this capacity confound. I considered the abstract/body discrepancy of 4.76% vs 4.67% in the online A/B report and the Section 5.2 polynomial proof; the former is a reporting inconsistency that can be corrected by checking the A/B logs, and the latter's degree recursion is obscured by missing superscripts but can be repaired, so neither is as load-bearing as the parameter mismatch. Until a parameter-matched HSTU baseline is run, the architectural attribution remains conditional, so the reader's CONDITIONAL verdict is appropriate.","tokens_in":19386,"tokens_out":8207,"duration_ms":70457,"concrete_test":"Report exact non-embedding parameter counts for every row of Table 2. Then train an HSTU-Large baseline whose total non-embedding parameter count is matched to FuXi-alpha-Large (e.g., by increasing HSTU hidden dimension or adding FFN stages) on MovieLens-20M and KuaiRand, using the same sampled-softmax loss, optimizer schedule, and compute budget. If the matched HSTU closes the NDCG@10 gap in Table 2, the FuXi-alpha advantage is capacity-driven; if HSTU-matched remains clearly below, the architectural claim survives. Separately, parameter-match the Table 5 ablations by adding an MLP of equivalent width to w/o-AMS and w/o-MFFN variants.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that FuXi-alpha's architecture, rather than extra scale, drives its superiority over HSTU. Section 6.1.4 states 'we maintain the same model parameters as HSTU in the first two datasets, except for the number of layers,' but this is not a total-parameter match. Section 4.2.1 adds per-layer temporal/positional bucket embeddings and a projection matrix W_u in the AMS; Section 4.2.2 adds a MFFN with W_o and SwiGLU projections W1-W3; and Section 5.4.2 states that HSTU 'lacks an FFN layer.' At identical hidden dimensions and layer counts, FuXi-alpha therefore has strictly more non-embedding parameters per layer than HSTU. Table 2 and the layer-scaling figures cannot separate architectural benefit from capacity, so the claimed 'consistently obtains the best results' and the conclusion that performance 'continually improves' as layers grow are confounded. The ablations in Table 5 are also not parameter-matched, so they do not resolve the attribution.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces FuXi-alpha, an autoregressive sequential recommendation transformer whose FuXi block replaces standard self-attention with an Adaptive Multi-channel Self-attention (AMS) mechanism that routes semantic, temporal, and positional signals through separate channels, and replaces the usual FFN with a two-stage Multi-stage FFN (MFFN) for implicit feature interactions. The authors compare against BPRMF, GRU4Rec, NARM, SASRec, LLaMa, and HSTU on MovieLens-1M, MovieLens-20M, KuaiRand, and a private Huawei Music dataset; report consistent wins across all public datasets and layer depths; report scaling behavior with increasing layer count; and report a 7-day online A/B test in Huawei Music with a 4.76%/4.67% plays-per-user increase and a 5.10% listening-duration increase. Code is released.","tokens_in":19551,"tokens_out":8184,"duration_ms":62485,"significance":"If the empirical claims hold, FuXi-alpha is a strong candidate for industrial sequential recommendation, and the paper provides useful evidence that autoregressive generative recommendation can be improved by more explicit temporal/positional modeling plus additional FFN capacity. The released code, public-benchmark results, and the held-out industrial traffic evaluation are concrete strengths. However, the core attribution claim, that the architecture rather than additional parameters drives the gains, is not established by the reported experiments, and the scaling-law statement is not supported by a fitted law. These issues materially affect the central contribution, so the paper needs a major revision before the claims can be accepted.","major_comments":[{"comment":"The comparison with HSTU is not parameter-matched. The paper states that 'we maintain the same model parameters as HSTU in the first two datasets, except for the number of layers' (§6.1.4), but FuXi-alpha's per-layer parameter count is strictly larger than HSTU's: the AMS adds the projection W_u ∈ R^{d×3d_h} in Eq. (4) plus per-layer relative position/timestamp bucket embeddings, the MFFN adds W_o in Eq. (5) and the SwiGLU matrices W_1, W_2, W_3 in Eq. (7), and §5.4.2 notes that HSTU lacks an FFN layer. Consequently Table 2 and the layer-scaling curves cannot separate the benefit of the AMS/MFFN design from the added capacity, and the ablations in Table 5 have the same confound. Please add parameter-matched comparisons (e.g., matching total non-embedding parameters by adjusting hidden sizes or by adding a comparable FFN to HSTU), or explicitly reposition the results as a capacity-uncontrolled comparison and soften the attribution.","section":"§6.1.4, §4.2.1–4.2.2, Table 2"},{"comment":"The claim that FuXi-alpha 'adheres to the scaling law' is not supported by the evidence. Figure 5 shows only a monotone trend over layer counts 2–32 on the industrial dataset with no fitted power law, no held-out scaling-law analysis in terms of data or compute, and no statistical uncertainty; Figure 6 adds only four points per dataset. More importantly, Figure 6(a) shows that on MovieLens-1M NDCG@10 and HR@10 decline when going from 8 to 16 layers, which directly contradicts the conclusion in §7 that 'the performance continually improves while scaling up our model.' Please either fit and report a specific scaling relation or explicitly state the weaker observed trend, and discuss the MovieLens-1M failure case.","section":"§6.2.3, §7, Figure 6"},{"comment":"The polynomial-approximation argument as written is not correct. After substituting the inductive hypothesis into Eq. (10), the expression Σ_j a_{i,j}^{(l+1)} x_{0,j} F_{2^l−1} + 1 has degree at most 2^l, so the claim that every term with total degree up to 2^{l+1} appears in this expression is wrong; the stated equality to F_{2^l} in Eq. (12) needs a degree bound of 2^l, and the subsequent step to degree 2^{l+1}−1 requires a separate argument about the product with the outer F_{2^l−1}. Please correct the induction indexing or state a more limited 'degree at most' result. The proof also relies on the simplified block in Eq. (9), so the conclusion should be framed as an expressiveness statement for the simplified model rather than an exact characterization of the full FuXi block.","section":"§5.2, Eqs. (10)–(12)"}],"minor_comments":[{"comment":"The reported online A/B play-count lift is inconsistent: the Abstract and §1 report 4.76%, while §6.6 reports 4.67% for the same metric.","section":"Abstract vs. §6.6"},{"comment":"The text says the deep variants are denoted 'XX-Large' after stacking 4x layers to 8 layers, but Table 2 and the surrounding text use the suffix '-Large'; please align the naming.","section":"§6.1.4, Table 2"},{"comment":"The variant label 'w/o MFFN' conflicts with the body text, which says that removing only the second stage of the MFFN causes the drop; renaming the variant to 'w/o MFFN-stage-2' would make the ablation unambiguous.","section":"§6.4, Table 5"},{"comment":"The scaling, hyperparameter, and efficiency results report single runs without error bars or variance information; adding standard deviations or at least stating the number of seeds would improve reliability.","section":"Figures 5–8, Table 4"},{"comment":"There are minor language errors, for example 'we process them using a similar manner to HSTU by ourself' and 'users in this platform are usualy very active'; these should be copyedited.","section":"§6.1.1"},{"comment":"The notation in Eq. (16) uses W_o1 and W_o2 as separate output projections for the semantic and positional channels, whereas Eq. (5) defines a single W_o; please clarify whether these are the same weights or a notational simplification.","section":"§5.3, Eq. (16)"}],"recommendation":"major_revision","confidential_remarks":"The main concern is the parameter-match confound, which affects the central architectural-attribution claim and the scaling interpretation. The paper is otherwise suitable for a conference venue, and the code release plus industrial A/B test are assets. If the authors add matched-capacity comparisons and correct the scaling-law and Section 5.2 statements, the contribution would be solid."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"FuXi-alpha is a well-engineered sequential recommender. The core idea—decoupling semantic, temporal, and positional channels in self-attention, then adding a two-stage FFN for implicit interactions—is simple and effective. The offline gains over HSTU are consistent across MovieLens-1M, MovieLens-20M, and KuaiRand, and the ablations show both AMS and MFFN contribute. They also release code and report a live A/B test on Huawei Music. That is real evidence and should be credited.\n\nThe main soft spot is the comparison with HSTU. Section 6.1.4 says they keep the same model parameters as HSTU, but that is not a total-parameter match: AMS adds temporal/positional embeddings and a projection, and MFFN adds a whole extra FFN block. So Table 2 and the scaling figures cannot cleanly separate architectural benefit from added capacity. The ablations have the same issue. This is not fatal, but it means the central attribution claim is weaker than the paper suggests. A parameter-matched baseline, or at least a table of parameter counts, would fix it.\n\nThe scaling-law claim is also loose. They show monotonic improvement with layers on the industrial dataset, but there is no fitted power law, and on MovieLens-1M performance degrades at 16 layers. The conclusion says performance \"continually improves\" which is only true for the industrial setting. The A/B numbers also have a typo: 4.76% in the abstract versus 4.67% in Section 6.6. Minor, but embarrassing. The polynomial proof in Section 5.2 is a bit sloppy in the induction step—the claimed degree range inside the bracket is off by one—but the overall degree bound appears correct for the simplified block they analyze.\n\nNone of these are deal-breakers. The architecture is new as a combination, the empirical work is broad, and the deployment result gives practical credibility. With a parameter-matched comparison and cleaned-up scaling and A/B statements, this would be a solid contribution. I would send it to peer review and expect the reviewers to ask for exactly those revisions.\n\nFor you: worth a read if you care about sequential recommendation or scaling claims in recsys. I would probably cite it as a strong HSTU successor, but not before the parameter question is sorted.","headline":"FuXi-alpha is a solid sequential-recommendation paper with consistent wins over HSTU and a real deployment, but the headline comparison is not parameter-matched and the scaling-law claim is overstated.","tokens_in":20138,"tokens_out":4986,"would_cite":true,"duration_ms":38169,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims that FuXi-α, an autoregressive recommender whose attention is split into semantic, temporal, and positional channels, beats all tested baselines and keeps improving with scale.","keywords":["sequential recommendation","generative recommendation","feature interaction","scaling laws","adaptive multi-channel self-attention","temporal information","positional information","online A/B testing"],"falsifier":"Train FuXi-α and HSTU with identical total parameter counts and identical FLOPs—not just identical listed hyperparameters—on MovieLens-20M and KuaiRand and compare NDCG@10; if the margin collapses, the central architectural claim fails.","tokens_in":19118,"feed_emoji":"🎵","tokens_out":10757,"duration_ms":85983,"temperature":0.7,"pith_summary":"FuXi-α is an autoregressive sequential recommendation model built on the idea that temporal and positional signals deserve their own attention channels rather than being added into semantic attention weights or blended into latent vectors. The paper's central claim is that this decoupling, together with a two-stage feed-forward network for implicit feature interactions, produces consistent gains: FuXi-α reports the best HR@10/50, NDCG@10/50, and MRR on MovieLens-1M, MovieLens-20M, and KuaiRand against SASRec, LLaMa, HSTU, and older baselines. If true, the result matters because it offers a concrete architectural recipe for scaling sequential recommenders along LLM-style scaling laws, and because a deployed version in Huawei Music raised songs played per user by 4.76% and listening time by 5.10%. The paper frames the model as restoring the balance between explicit and implicit feature interactions that HSTU tipped too far toward the explicit side.","feed_headline":"Beats HSTU on all public benchmarks via three-channel attention","feed_subtitle":"Splitting attention into semantic, temporal, and positional signals lifts offline metrics and Huawei Music plays by 4.76%.","key_machinery":"The load-bearing mechanism is the Adaptive Multi-channel Self-attention (AMS) layer together with the Multi-stage Feed-Forward Network (MFFN). AMS runs three parallel attention channels—semantic (learned query/key/value attention), temporal (bucketed timestamp differences with per-bucket learnable weights), and positional (learned relative-position weights)—shares the value vectors across channels, concatenates the channel outputs, normalizes them, and gates the result by element-wise multiplication with a projection of the layer input. The MFFN first fuses the channel outputs with the layer input and then applies a SwiGLU feed-forward stage for implicit interactions. The formal account is a polynomial-approximation induction: after $b$ stacked FuXi blocks, the output at a position is $x_{0,i} F_{2^b-1}$ with interaction degree up to $2^b-1$, showing why added depth adds expressive power.","core_discovery":"FuXi-α's central discovery is that sequential recommendation improves when self-attention stops folding time and position into semantic attention weights and instead runs three parallel channels—semantic, temporal, and positional—that share value vectors, then fuses them with a two-stage feed-forward network. The paper argues that this restores the implicit feature interactions that HSTU de-emphasized and makes temporal and positional cues more expressive. After $b$ stacked blocks the output is $x_{0,i} F_{2^b-1}$, an interaction polynomial of degree up to $2^b - 1$, which the paper offers as the formal reason performance keeps climbing with depth. In offline evaluation FuXi-α reports the best HR@K, NDCG@K, and MRR on MovieLens-1M, MovieLens-20M, and KuaiRand, and in an online A/B test it reports a 4.76% increase in songs played per user and a 5.10% increase in listening duration.","pith_inferences":["If the channel separation, not the extra parameters, drives the gains, the AMS design should transfer to other Transformer-based sequence models where timestamp and position are side information.","Varying the number of temporal buckets while holding everything else fixed would show how much of the gain depends on time quantization, a lever the paper leaves untested.","The polynomial-approximation argument implies scaling gains should saturate once a dataset's required interaction degree is reached; MovieLens-1M already declines at 16 layers, so locating that saturation point on industrial data would test the mechanism's own account.","Because negative sampling showed larger returns than depth, the reported scaling law may be confounded with sampling pressure; a joint sweep of data size, negatives, and layers would separate the two."],"forward_implications":["At two layers FuXi-α beats the strongest baseline by 13.24% in NDCG@10 averaged over the three public datasets; at eight layers the margin over HSTU-Large is 7.26% in NDCG@10.","On the industrial music dataset, scaling from 2 to 32 layers improves NDCG@10 and HR@10 monotonically, which the paper reads as adherence to a scaling law.","Adding negative samples from 32 to 256 produces larger gains than increasing layers from 2 to 16 on the tested datasets, making sampling ratio a first-order scaling knob.","A 7-day A/B test in Huawei Music reports 4.76% more songs played per user and 5.10% more listening time per user relative to a mature multi-channel baseline, after which the model became a permanent channel.","FuXi-α achieves throughput comparable to HSTU but lower than SASRec and LLaMa, so the performance gains come with a measurable speed cost."],"supporting_citations":[{"why":"Defines the HSTU baseline, supplies the preprocessed MovieLens train/validation/test sets, and sets the parameter and optimizer configuration that FuXi-α is matched against.","marker":"[72]"},{"why":"Introduces SASRec, the foundational self-attentive sequential recommender and the strongest conventional baseline in the comparison.","marker":"[23]"},{"why":"Supplies the LLaMa baseline and the RMSNorm-plus-SwiGLU feed-forward design that the MFFN builds on.","marker":"[10]"},{"why":"Provides the bucketed relative-position and timestamp embedding scheme that the AMS temporal and positional channels adapt.","marker":"[42]"},{"why":"Provides root mean square layer normalization used inside the AMS and MFFN blocks.","marker":"[75]"},{"why":"Justifies the sampled softmax loss with random negatives used to train and scale the model efficiently.","marker":"[25]"},{"why":"Establishes the LLM scaling-law background that motivates treating layer count as a scaling axis for recommenders.","marker":"[24]"}],"fun_headline_variants":["Three-channel attention boosts recommendations and Huawei Music plays","FuXi-α: separate semantic, temporal, positional attention channels","Multi-channel attention fuses semantic, temporal, positional signals","FuXi-α lifts Huawei Music plays 4.76%, listening 5.10%","Feature-interaction transformer with adaptive multi-channel attention"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The comparisons are read as matching HSTU's capacity because the listed parameter settings match, even though the added attention channels add weights that are not counted in that match, so some of the reported gain could be capacity rather than the mechanism.","fun_headline_variants_meta":{"raw":{"variants":["Three-channel attention boosts recommendations and Huawei Music plays","FuXi-α: separate semantic, temporal, positional attention channels","Multi-channel attention fuses semantic, temporal, positional signals","FuXi-α lifts Huawei Music plays 4.76%, listening 5.10%","Feature-interaction transformer with adaptive multi-channel attention"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000992,"raw_usage":{"total_tokens":4236,"prompt_tokens":1008,"completion_tokens":3228,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":624,"completion_tokens_details":{"reasoning_tokens":3141}},"tokens_in":624,"tokens_out":3228,"duration_ms":21671,"temperature":1.0,"reasoning_tokens":3141,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-09T10:06:07.317266+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train FuXi-α and HSTU with identical total parameter counts and identical FLOPs—not just identical listed hyperparameters—on MovieLens-20M and KuaiRand and compare NDCG@10; if the margin collapses, the central architectural claim fails.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Justifies the sampled softmax loss with random negatives used to train and scale the model efficiently."}],"review_version":1}